aaron/ispc
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

loop execution for i8/16 32-addr-bit gathers/scatters

Browse Source
This commit is contained in:
Anton Mitrokhin
2015-03-26 14:53:19 +03:00
parent a5b2695771
commit 81cb374084
1 changed files with 30 additions and 31 deletions
Download Patch File Download Diff File Expand all files Collapse all files

53
examples/intrinsics/knl.h
View File

@@ -2789,27 +2789,27 @@ template <int ALIGN> static FORCEINLINE void __store(__vec16_i8 *p, __vec16_i8 v
*p = v; *p = v;
} }
static FORCEINLINE void static FORCEINLINE void __scatter_base_offsets32_i8(uint8_t *b, uint32_t scale, __vec16_i32 offsets,
__scatter_base_offsets32_i8(uint8_t *b, uint32_t scale, __vec16_i32 offsets,
__vec16_i8 val, __vec16_i1 mask) __vec16_i8 val, __vec16_i1 mask)
{ {
__vec16_i32 tmp = _mm512_extload_epi32(&val,_MM_UPCONV_EPI32_SINT8, // TODO
_MM_BROADCAST32_NONE, _MM_HINT_NONE); for (int i = 0; i < 16; ++i)
_mm512_mask_i32extscatter_epi32(b, mask, offsets, tmp, if ((mask & (1 << i)) != 0) {
_MM_DOWNCONV_EPI32_SINT8, scale, int8_t *ptr = (int8_t *)(_base + scale * offsets[i]);
_MM_HINT_NONE); *ptr = val[i];
}
} }
static FORCEINLINE void static FORCEINLINE void __scatter_base_offsets32_i16(uint8_t *b, uint32_t scale, __vec16_i32 offsets,
__scatter_base_offsets32_i16(uint8_t *b, uint32_t scale, __vec16_i32 offsets,
__vec16_i16 val, __vec16_i1 mask) __vec16_i16 val, __vec16_i1 mask)
{ {
__vec16_i32 tmp = _mm512_extload_epi32(&val,_MM_UPCONV_EPI32_SINT16, // TODO
_MM_BROADCAST32_NONE, _MM_HINT_NONE); for (int i = 0; i < 16; ++i)
_mm512_mask_i32extscatter_epi32(b, mask, offsets, tmp, if ((mask & (1 << i)) != 0) {
_MM_DOWNCONV_EPI32_SINT16, scale, int16_t *ptr = (int16_t *)(_base + scale * offsets[i]);
_MM_HINT_NONE); *ptr = val[i];
}
} }
@@ -2900,27 +2900,26 @@ static FORCEINLINE void __masked_store_blend_float(void *p, __vec16_f val,
static FORCEINLINE __vec16_i8 static FORCEINLINE __vec16_i8
__gather_base_offsets32_i8(uint8_t *base, uint32_t scale, __vec16_i32 offsets, __gather_base_offsets32_i8(uint8_t *base, uint32_t scale, __vec16_i32 offsets,
__vec16_i1 mask) { // TODO
// (iw): need to temporarily store as int because gathers can only return ints.
__vec16_i32 tmp = _mm512_mask_i32extgather_epi32(_mm512_undefined_epi32(), mask, offsets, base,
_MM_UPCONV_EPI32_SINT8, scale,
_MM_HINT_NONE);
// now, downconverting to chars into temporary char vector
__vec16_i8 ret; __vec16_i8 ret;
_mm512_extstore_epi32(ret.v,tmp,_MM_DOWNCONV_EPI32_SINT8,_MM_HINT_NONE); for (int i = 0; i < 16; ++i)
if ((mask & (1 << i)) != 0) {
int8_t *ptr = (int8_t *)(_base + scale * offsets[i]);
ret[i] = *ptr;
}
return ret; return ret;
} }
static FORCEINLINE __vec16_i16 static FORCEINLINE __vec16_i16
__gather_base_offsets32_i16(uint8_t *base, uint32_t scale, __vec16_i32 offsets, __gather_base_offsets32_i16(uint8_t *base, uint32_t scale, __vec16_i32 offsets,
__vec16_i1 mask) { __vec16_i1 mask) {
// (iw): need to temporarily store as int because gathers can only return ints. // TODO
__vec16_i32 tmp = _mm512_mask_i32extgather_epi32(_mm512_undefined_epi32(), mask, offsets, base,
_MM_UPCONV_EPI32_SINT16, scale,
_MM_HINT_NONE);
// now, downconverting to chars into temporary char vector
__vec16_i16 ret; __vec16_i16 ret;
_mm512_extstore_epi32(ret.v,tmp,_MM_DOWNCONV_EPI32_SINT16,_MM_HINT_NONE); for (int i = 0; i < 16; ++i)
if ((mask & (1 << i)) != 0) {
int16_t *ptr = (int16_t *)(_base + scale * offsets[i]);
ret[i] = *ptr;
}
return ret; return ret;
} }